PeopleMatters is an HR publication that reviews and links HR products. Octopus is their on-site AI search — users ask questions like “what’s best for time-tracking at a mid-sized company” and it recommends products from PeopleMatters’ indexed catalog.
The shipped version was a thin LLM-wrapped keyword search: convert query → keywords → top-ranking products → return them. Mechanical, brittle, and bad at any question that wasn’t already shaped like a product filter.
What I built
An agentic version closer in spirit to my Founders Sage replica. Instead of converting the query into keywords, the agent browses contextual information about products — what they do, who they’re for, the trade-offs — and reasons about fit. A query about a mid-sized company gets a different answer to the same query about a 50-person startup, because the agent is actually thinking about the situation rather than ranking by keyword overlap.
Outcome
Never replaced production. From what I remember, it didn’t fit the funding/structure on the company side and a working version already existed, so the rebuild slid down the priority list and was never picked up again.